[ptx] Re: fine tuning 2

[Pablo d'Angelo]

On Sun, 11 Jul 2004, Ph. Gac wrote:

> 
> hi Pablo,
> 
> thanks for your answer and thanks a lot for your software which I am now using instead of PTOpenGUI since it is quite
> stable.
> 
> I confirm, with your last release, that when you change X in the text box, and then select Y with the mouse or with TAB,
> all the textboxes get cleared and the CP pair is deselected; it cannot be selected again by clicking or double-clicking
> in its row, so you have to select another pair and then select it again to change Y!

hmm, strange. looks like another difference between wxGTK and wxMSW.

> I'm working on Windows 2000 SP4 with GTK 2.2.4 and also GTK 1.3 which I need for GIMP 1.25 (GIMP 2.0 is really too buggy
> for now). It's running on an Athlon XP 1500 with 512 MB DDRam and a 1600x1200x65k screen ("high color", large fonts, but
> no display problem so far). It's fine for Hugin but for the one pixel wide features which are sometimes hardly
> noticeable.

hugin doesn't use GTK on windows, it uses the native MFC controls.

Do you really see the need to select control points with a 16x
detailed view? I'm a bit sceptic that it improves results so much.

I'll add a subpixel finetune later on, that will also work on rotated
views.

However I don't think that most people want to work with subpixel control
point choosers. I'll rather work on a fine tune that works better with wide
angle or fisheye images.

> I began with PTOpenGUI which has a 16x detailed view and allows to set fractional coordinates. These views are generated
> by interpolation and it seems they are also warped to help the feature identification. Then I could achieve, in some
> specific cases, a 0.5 or even (?!) 0.2 pointing precision. This happened to be useful and I was quite surprised.

Do you have a project where this effect can be seen? I don't really belive
in it. I'd rather add 2 or 3 more auto-finetune points.

> Thinking about it, 3 ideas emerged...
> 1- To infer the perspective parameters, one can in theory use any couples of overlapping points but the ones with the
> biggest homographic warping give more significant results. The trouble is that these points also get the biggest optic
> distortion, some of which is not taken into account by the a-e parameters (maybe, especially with aspherical lens).

Well, after the correction with a-e, it should just be the pinhole camera
projection (hopefully that is..). Currently the fine tune matcher assumes an
orthographic projection, but that can be accepted with pinhole projection
images, the usual 40° Hfov or so, I think, especially if features with a
point or some other shape that is not changed too much by the different
projection.

Its completely broken for fisheye and wideangle images.

> 2- a small error in the first image of the series may induce larger and larger errors to adapt consecutives views which
> results in big discrepancies when you want to make the ends meet. Then, global optimization sometimes gets crazy,
> especially if you are tempted to let the FOV vary...

usually one is doing global optimisation, execpt when the pairwise
optimisation of hugin is used. Since the optimisation is nonlinear, it might
find local minima, especially if the solution is far from the starting
positions.

> 3- top-border points are often in the sky... bottom-border points are often on the foreground, and I seldom take my
> 1.5kg 30cm pod with my 200g 8cm camera... so there are sometimes few usable points. This low number of CPs doesn't allow
> the compensation of errors so the coordinates should be as precise as possible.

In that case you might have to use some more overlap to specify more points.

> I dont use autopano because it is unusable with modern buildings I'm working on now, which have lots of repeated
> features... I used it with a mountain panorama and it gave also poor results, choosing points in distorted areas, or on
> the foreground. Hand tuning with PTOpenGUI was very slow but gave a very good compromise. Second best choice was Hugin
> autofinetuning.

Hmm, I usually get very good results with the autopano-sift package, but I
haven't tried the typical city pano with many similar windows etc.

ciao
  Pablo